Expand description
Online learning and dynamic retraining infrastructure Online Learning Infrastructure for Dynamic Model Retraining
This module provides incremental model updates without full retraining, supporting continuous improvement in production ML systems.
§References
- [Bottou 2010] “Large-Scale Machine Learning with Stochastic Gradient Descent”
- [Crammer et al. 2006] “Online Passive-Aggressive Algorithms”
§Toyota Way Principles
- Kaizen: Continuous model improvement via online learning
- Jidoka: Drift detection stops bad predictions automatically
- Just-in-Time: Retrain only when drift detected, not on schedule
Modules§
- corpus
- Corpus Management for Online Learning
- cpt
- Continual Pre-Training (CPT) Pipeline (GH-448)
- curriculum
- Curriculum Learning for Progressive Training
- dam
- Differentiable Adaptive Merging (DAM) (GH-446)
- distillation
- Knowledge Distillation for Model Compression
- distillation_
advanced - Advanced Knowledge Distillation Strategies (GH-451)
- dpo
- Direct Preference Optimization (DPO) (GH-449)
- drift
- Drift Detection for Triggering Model Retraining
- eval_
harness - Evaluation Harness for Standard Benchmarks (GH-454)
- moe_
construction - Mixture of Experts (MoE) Construction from Dense Models (GH-445)
- orchestrator
- Retrain Orchestrator for Drift-Triggered Model Updates
- per_
layer_ merge - Per-Layer Merge Granularity for Model Merging (GH-452)
- rlvr
- Reinforcement Learning on Verifiable Rewards (RLVR) (GH-450)
- tokenizer_
surgery - Tokenizer Surgery for Vocabulary Transplantation (GH-447)
Structs§
- Online
Learner Config - Configuration for online learning
- Online
Linear Regression - Simple online linear regression using SGD
- Online
Logistic Regression - Simple online logistic regression using SGD
Enums§
- Learning
Rate Decay - Learning rate decay schedules
Traits§
- Online
Learner - Online learning capability for incremental model updates
- Passive
Aggressive - Passive-Aggressive online learning for classification